光伏发电作为缓解当前能源危机的重要手段之一，具有绿色低碳、安全可靠、分布灵活等优点，已备受关注。然而，光伏组件在长期使用过程中易被尘土、落叶等遮挡产生热斑故障，严重影响光伏组件使用寿命及工作可靠性。因此，本文以某光伏电站提供的无人机巡检数据集作为研究对象，开展了基于角度感知和模型迁移的光伏组件热斑故障检测方法研究。该研究对于提高光伏电站巡检效率及实现光伏电站运维智能化具有重要意义。

论文主要研究内容如下：

（1）针对传统算法实时性不足，且因热斑目标像素占比低而导致故障特征难以有效表达，进而限制算法检测性能的问题，提出一种知识蒸馏机制下双分支协同训练的光伏热斑故障检测算法。首先，设计“教师+学生”协同训练架构以借助教师深层检测网络优势提升算法检测精度，并结合学生网络小参数量特点提高算法推理效率；其次，选用YOLOX检测算法作为学生网络，通过构建无锚框模型来解决先验锚点框对小尺寸目标适应性较弱的问题；然后，在教师网络中构建CSPHN模块以捕捉故障目标高阶特征交互信息；接着，设计双分支多级特征自适应融合模块以并行融合方式聚合全局及局部多级特征；最后，提出可变形上下文Transformer模块，通过在多头自注意力机制中构建偏移网络来挖掘特征图中丰富的动静态上下文信息。实验结果表明，所提出算法能够在不增加额外计算成本的条件下，实现密集小尺度热斑故障目标的准确识别任务，检测精度达到82.1%，相较于原网络提升了1.4%。

（2）针对复杂环境下因待检测目标存在特征模糊、角度畸变等特性导致水平框检测算法难以准确定位热斑故障的问题，提出一种基于旋转框精细定位的分割-检测串行光伏热斑故障检测算法。首先，设计改进DeepLabv3+分割网络从伪高亮背景区域中捕捉精细光伏板边界特征；接着，构建交叉自适应频率Transformer模块，借助多头频率交叉自注意力机制充分挖掘多尺度特征之间的信息交互特性；然后，针对颈部网络难以捕获非相邻特征图之间相互作用的问题，提出跨尺度移位网络以提升检测网络对跨尺度特征的远距离交互能力；最后，引入凸包特征自适应优化算法，利用旋转框思想准确定位角度各异的热斑故障目标，并借助联合优化方式缓解多目标特征混叠现象。实验结果表明，相较于11种经典检测网络，所提出算法能够在各类复杂环境下准确反映热斑目标真实形状，检测精度高达92.3%。

As one of the important means to alleviate the current energy crisis, photovoltaic power generation has attracted much attention due to its advantages such as green and low-carbon, safe and reliable, and flexible distribution. However, during the long-term operation of photovoltaic modules, they are prone to be covered by dust and fallen leaves, resulting in hot-spot faults that seriously affect the service life and operational reliability of photovoltaic modules. Therefore, this paper takes the unmanned aerial vehicle inspection data set provided by a photovoltaic power station as the research object, and conducts research on the hot-spot fault detection method of photovoltaic modules based on angle perception and model transfer. This research is of great significance in improving the inspection efficiency of photovoltaic power stations and achieving intelligent operation and maintenance of photovoltaic power stations.

The specific contents are as follows:

(1) To address the problem that traditional algorithms have insufficient real-time performance and the fault features are difficult to effectively express due to the low proportion of hot-spot target pixels, which limits the detection performance of the algorithm, a bi-branch collaborative training algorithm based on knowledge distillation mechanism for hot-spot fault detection is proposed. Firstly, a "teacher+student" collaborative training architecture is designed. Such mechanism takes advantage of the deep detection network of a teacher model to enhance detection accuracy and combines the characteristics of the small parametric student network to improve the efficiency of the algorithm inference. Secondly, a YOLOX detection algorithm is adopted as the student network, and an anchor-free model is constructed to solve the problem of weak adaptation of prior anchor boxes to small-sized targets. Then, a CSPHN module is constructed in the teacher network to capture high-order feature interaction information of faulty targets. Next, a bi-branch multi-level feature adaptive fusion module is designed to aggregate global and local multi-level features in a parallel fusion manner. Finally, a deformable context Transformer module is proposed to mine the rich dynamic and static context information by constructing an offset network in a multi-head self-attention mechanism. Experimental results show that the proposed algorithm can accurately identify densely distributed small-scale hot spot fault targets without increasing additional computing costs, with a detection accuracy of 82.1%, which is 1.4% higher than that of the original network.

(2) In response to the challenge that the horizontal box detection algorithm is difficult to accurately locate hot-spot faults due to the characteristics of feature blurring and angle distortion of the fault targets under complex environments, this paper proposes a segmentation-detection serial photovoltaic hot-spot fault detection algorithm based on refined rotated boxes. Firstly, an improved DeepLabv3+ segmentation network is designed to capture fine photovoltaic panel boundary features from pseudo-highlight background areas. Then, a cross-adaptive frequency Transformer module is constructed to fully exploit the information interaction characteristics between multi-scale features by leveraging the multi-head frequency cross-self-attention mechanism. Next, to address the problem of the neck network's difficulty in capturing interactions between non-adjacent feature maps, a cross-scale shift module is proposed to enhance the detection network's long-distance interaction ability with cross-scale features. Finally, the convex hull feature adaptive optimization algorithm is introduced to accurately locate hot-spot faults with different angles by using the idea of rotation box, and the joint optimization method is used to alleviate the phenomenon of multiple target feature overlapping. Experimental results demonstrate that compared to 11 classic detection networks, the proposed algorithm can accurately reflect the real shape of hot-spot faults in various complex environments, with a detection accuracy of up to 92.3%.